Hi David,

Good job! If I can find the time I'll try to rerun your tests to see if I get the same results. Some questions/comments:

You state in the overview table that you use a Java 6 version of JRockit, but later in the text that you use a 5.0 version. Which is correct?

I'm a bit suprised about the xmlmark results. In my earlier test on a 2.67 GHz Clovertown I got ~1400 base and ~1750 peak on SAX1; a 25% benefit from tuning compared to your 10%. However, my config did differ a bit from yours: I used a 32-bit OS and JVM and a smaller heap (1.5GB), and I did not use -XXtlasize for peak, and ran with HW prefetching disabled in the BIOS. It is possible that you have hit a memory bus bottleneck which limits the benefit from tuning.

You will be able to see the IPC benefits of Harpertown on jbb when Intel announces the launch results (on Monday?). As a baseline, a fully tuned 3.0 GHz Clovertown tops out around 250k with a 4-JVM config, BIOS tuning and the latest release of JRockit, see:
http://www.spec.org/jbb2005/results/res2007q4/jbb2005-20071023-00391.html

Recent SPEC publications almost invariably use multiple JVMs. I like your approach to & reason for going with one JVM, but note that you may have run into a GC bottleneck with high alloc rates and a "small" heap.

Agree that jbb is a little understood benchmark. Maybe it's time to remedy that and write an article about it? I can probably help if you want to pursue this.

Cheers,

Henrik